MiniMax H3 Open-Weights Day-Zero: An Omni-Modal Video Model That Runs on a 3060

Hi guys, Mr. Technology here.

MiniMax shipped H3 today — their third-generation video model and the first released with open weights. It handles text-to-video, image-to-video, first-and-last-frame, and reference-to-video (up to 12 mixed-modality files: text, images, video clips, audio clips). Output is 2K resolution, up to 15 seconds, 24 FPS, with native stereo audio generated in the same pass as the video. ComfyUI shipped day-zero support. Here is what shipped, what is actually new, and why this release matters more than the launch headlines suggest.

What Actually Shipped

H3 is the third generation of the Hailuo line. The headline architectural leap is not the 2K or the 15 seconds — it is that audio is now a first-class output of the same forward pass as video, not a post-process bolted onto rendered frames.

Four input modes: T2V (prompt only), I2V (single image to life), FL2V (first-and-last-frame, opening and closing controls), and R2V (reference-to-video, up to 9 images, 3 video clips, or 3 audio clips as cross-modal references). Eleven languages get stable dialogue support out of the box (Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish). Aspect ratios cover 21:9, 16:9, 4:3, 1:1, 3:4, 9:16. Output audio is 32 kHz stereo, native to the generation pass.

The Architecture

Three modules, two open:

1. H3-Context-IR (closed): a hosted preprocessor that parses free-form multimodal inputs into a structured "Context Intermediate Representation." MiniMax exposes an API and prompting guidance, but the model itself is not in the open release. 2. H3-Base (open): the generator. 768p video with native stereo audio. 3. H3-Regenerate-2K (open): re-runs generation at 2K, conditioned on the 768p result plus the original context. This is a re-generation, not an upscale — it pulls sharper detail from the original inputs.

H3-Base is clean two-stream design: a full pretrained Qwen3-VL-32B encoder (50th-layer hidden states feed the generator) for text and visual inputs; H3-VisualVAE (16× spatial, 4× temporal, 24 channels, temporally causal, 1×2×2 patchified for 32× spatial downsample at the token level); H3-AudioVAE (32 kHz stereo, 40 Hz latent rate, independent channel encoding); and a 33B-parameter dense, single-stream H3-Omni-Transformer with RoPE for spatial-temporal relationships. The ~13B parameters in AdaLN branches can be precomputed and cached — they do not need to be loaded at inference.

The architecture natively supports sparse attention for training and inference, though the initial release ships full attention only. Sparse-attention inference is in a future update.

The Memory Math

This is the part that should change your build plans.

ComfyUI's team published the inference engineering work alongside the release:

• Full-precision footprint: 123.6 GB
• Smallest optimized variant: 42.5 GB
• 66% memory reduction

The technique is structural. The modulation weights in H3-Omni-Transformer are roughly 40% of total parameters. ComfyUI found these could be pruned and replaced with a functionally equivalent lookup table, with no measurable loss in output quality. On top of that, the weights ship with an accurate int8 convrot quantization, and custom kernels cut peak VRAM during inference.

The combined effect — plus ComfyUI's dynamic VRAM offloading — is that a 2K video model with native audio runs on an RTX 3060. That is a 12 GB consumer card. Not an A100, not a 4090, not even a 4090 Ti — a 3060.

If you are building local-first video pipelines, this is the first release where "runs on a workstation" and "2K with audio" are simultaneously true.

Pricing and Availability

Open weights on HuggingFace: **MiniMaxAI/MiniMax-H3 (original) and Comfy-Org/MiniMax-H3** (ComfyUI-repackaged with int8 convrot and pruned int8/fp8 variants ready to drop in). API: platform.minimax.io (global), platform.minimaxi.com (CN). Apps: hailuoai.video (web), hub.minimax.io (desktop).

ComfyUI 0.30.0 ships day-zero support with I2V, T2V, and R2V workflow templates. Comfy Cloud runs H3 on managed GPUs if you do not want to wire up the local stack.

License is the MiniMax H3 Community License — research and commercial use allowed with use-case restrictions. Read it before commercial deployment, especially if you plan to use H3 outputs as training data.

The closed Context-IR preprocessor means the full demo quality requires the hosted API. Plan to call the API for prompt parsing, then run the generator locally for cost.

Does It Matter?

Yes. Three reasons this release is structurally important, not just incrementally interesting.

1. Open weights + audio + 2K is a new bar. Wan 2.2, CogVideoX, and the open-weights video stacks of early 2026 all shipped either open weights or audio, not both. MiniMax shipped both, with quantization work that puts 2K on consumer hardware.

2. Cross-modal is no longer a research demo. H3 takes up to 12 mixed-modality files (text + images + video + audio) and resolves them against a prompt that describes their relationship. That is closer to how creative work actually happens than the standard T2V pipeline.

3. Native stereo audio in one pass is the unlock. Every previous "open video model with audio" shipped audio as a post-process. H3 generates video and stereo audio jointly from the same conditioning. Voice-over, music-driven video, and dialogue scenes no longer need a separate audio stage.

The "does it actually look as good as the demo" question is still open — I have not benchmarked it on production footage. But the architectural choices are sound, and the open-weights release makes independent evaluation possible, which is the precondition for any of this mattering in production.

What This Means For Engineers

• Video agents: H3 is the open-weights backbone to evaluate against. Wan 2.2 still wins on certain motion-consistency benchmarks; H3 wins on audio, 2K, and reference-driven generation. Plan a head-to-head before committing.
• Audio-conditioned video: The native stereo pipeline eliminates the audio post-process. That is the single biggest workflow simplification in the release.
• Local-first video: 12 GB consumer GPU plus 64 GB system RAM is enough for 2K with audio. That changes the cost model for productized video tools.
• Inference scaling: The AdaLN-caching pattern — precompute ~13B parameters once, do not load at inference — is the architectural idea worth stealing. Any AdaLN-heavy diffusion or flow-matching model can adopt the same trick to reduce serving footprint.
• Training data: H3 outputs can become training signal for downstream models under the community license. Read the terms before doing this commercially.

The smart move this week: pull the weights, run the T2V and I2V workflows in ComfyUI, benchmark on your own footage. The smart move this quarter: redesign your video pipeline around native audio instead of post-process audio. The smart move this year: assume every flagship video model released after August 2026 ships audio as a first-class output, and build your data contracts and evaluation harnesses accordingly.

H3 is not the end of the open-weights video model race. It is the starting gun.

— Mr. Technology

Model: MiniMax H3 (third-generation video model, first open-weights release) Lab: MiniMax Release date: August 3, 2026 Open weights: MiniMaxAI/MiniMax-H3 (HuggingFace), Comfy-Org/MiniMax-H3 (ComfyUI-repackaged) Architecture: 33B-parameter dense single-stream H3-Omni-Transformer + Qwen3-VL-32B encoder + separate visual and audio VAEs Resolution: 2K (via H3-Regenerate-2K), 768p direct from H3-Base Duration: 4–15 seconds Frame rate: 24 FPS Audio: 32 kHz native stereo, generated in same pass as video Dialogue languages: 11 stable (Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish) License: MiniMax H3 Community License (commercial use with restrictions) Inference footprint: 42.5 GB on the smallest optimized variant (down from 123.6 GB), runs on RTX 3060 with dynamic VRAM offloading Quantization: int8 convrot, custom kernels, modulation-weight pruning to lookup table ComfyUI support: day-zero in v0.30.0; I2V, T2V, R2V workflow templates shipped API: platform.minimax.io (global), platform.minimaxi.com (CN) Closed component: H3-Context-IR preprocessor (API only, not in open release)